🔍 Read the full analysis: A Loop Problem, An AI Editor: GPT-6 Astra Adjusts Model Weights on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A November 2 blog post by four researchers affiliated with MIT CSAIL reports that GPT-6 Astra edited two layers of the open-weight Qwen3.5-0.8B model to reduce repetitive reasoning. On MATH-500, the reported score rose from 7.9% to 48.1%, while answers hitting the 32,768-token limit fell from 93.0% to 33.1%. The work is an early, unreviewed report, and the improvement followed an iterative search with substantial testing—not a single automated edit.
Four researchers affiliated with MIT CSAIL say OpenAI’s GPT-6 Astra edited two layers of the open-weight model Qwen3.5-0.8B, reducing repetitive outputs and raising its reported MATH-500 score from 7.9% to 48.1%. The November 2 account describes a weight-editing method that used no gradient descent, but the result came after an iterative, 20.3-hour search, and 33.1% of edited-model answers still hit the 32,768-token limit.
The work appears in a blog post titled Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The post includes logs and describes eight experiments. The authors call the results “very early”; the post is not a peer-reviewed paper and, according to the source account, is not on arXiv. OpenAI is not listed as an author, and no OpenAI statement about the experiment was identified by the source.
For the Qwen math test, the researchers gave Astra a Python environment to inspect and modify the model’s weights. The setup was CPU-only, had no internet access and allowed at most 300 code runs. Astra could assess behavior through limited model forward passes, with each pass covering up to 512 problems. The researchers also supplied practice questions from GSM8K and MATH, while keeping the final benchmark questions from the model-editing process.
The reported edit removed a direction associated with looping from output columns in the MLP layers numbered 15 and 19. The researchers say that changed 131,072 weights, or 0.016% of the model’s 0.8 billion weights. The published model files reportedly differ from the original only in the two layers’ down-projection tensors. The approach did not use gradient descent, backward passes or fitting the weights to a loss, but it did involve repeated testing and code-based analysis.
A Repair, Not a Math Breakthrough
The results point to a possible way to alter the behavior of an open-weight model without retraining all of its parameters. If the method proves repeatable, developers may be able to diagnose a narrow failure and make a targeted change while leaving most weights untouched. For people deploying open models, the practical interest is in behavioral repair: reducing a costly failure mode, such as an answer that repeats until it exhausts its token budget.
The benchmark figures do not show that Qwen acquired general mathematical ability. The base model already had a documented tendency to enter thinking loops, and the edit primarily addressed that weakness. The researchers themselves caution that “Fixing loops, however, is not the same as making a model better at math.” The MATH-500 score rose alongside a large, but incomplete, reduction in answers reaching the limit. That distinction matters for users weighing whether an edited model is reliable enough for real tasks.
The process also has costs and limits. The successful run took 20.3 hours and required 50 forward passes, with up to 512 problems per pass, according to the source account. Smaller searches produced much smaller gains. And because a changed set of weights is a distinct model version, anyone using a patch would need to evaluate and document that version rather than assume the original model’s behavior still applies.
Products Worth Considering
Apple Mac mini Desktop Computer with M4 chip, 10‑core CPU and GPU, 24GB Unified Memory, 512GB SSD, Gigabit Ethernet
Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD, Silver
Apple MacBook Air 13-inch with M5 chip, 16GB memory, 1TB SSD, and 12MP Center Stage camera (Midnight)
As an affiliate, we earn on qualifying purchases.
How the Loop Was Traced
Qwen3.5-0.8B is a small open-weight model, and its model card warns that it can produce thinking loops. In the researchers’ initial practice sample, 85 of 96 answers reached the 32,768-token limit. That failure gives the experiment its narrower purpose: identify and reduce one specific behavior, rather than improve every aspect of reasoning.
The researchers report that Astra linked the loops to a direction written by the MLPs in layers 15 and 19, then subtracted that direction from selected output columns. The reported MATH-500 results were 7.9% for the base model, 12.3% after 10 passes, 14.1% after 25 passes and 48.1% after 50 passes. The share of answers reaching the token limit fell from 93.0% to 87.2%, 86.4% and 33.1%, respectively. These are figures from the blog post’s experiment, not independent replication.
The post’s other experiments included attempts to identify behaviors or repair a failing car-driving policy. The source account says one task—identifying which of two fine-tuned Gemma models had been trained to flatter users—failed. The uneven results reinforce that the post presents a set of exploratory demonstrations, not evidence that an AI editor can reliably infer and change arbitrary model behaviors.
““Fixing loops, however, is not the same as making a model better at math.””
— The researchers, in their November 2 blog post
Replication and Reliability Remain Open
The result has not been established through peer review or independent replication in the material cited here. It is not clear whether the same edit would work across different prompts, sampling settings or model versions, or whether it could create other unwanted changes. The reported benchmark gains also do not establish performance on real-world math tasks.
The available figures show that looping persisted: 33.1% of edited-model MATH-500 answers still reached the maximum token count. The source material reports results for GSM8K and IFEval as well, but the supplied account does not provide the full experimental detail needed to judge statistical variation, run-to-run consistency or broader effects. Those limits make it difficult to separate a robust repair from a result specific to this setup.
The wording “without training” also needs qualification. The process avoided gradient descent and loss-based weight fitting, but Astra still used repeated model evaluations, practice problems and an extended search. The source account says the benchmark tests were withheld, but the degree to which the search or evaluation setup could influence the reported outcomes remains a question for further testing.
Independent Tests Will Matter
The immediate next step is independent replication: testing the published edited weights against the original under the same conditions, then checking performance across additional prompts and model versions. Follow-up work would also need to report whether the loop reduction holds consistently and whether the weight change affects other behaviors. The researchers’ logs and model files offer material for that scrutiny, but the supplied source does not identify a scheduled peer-reviewed paper or a planned replication.
For now, the report supports a limited conclusion: an AI system was used to inspect and modify a small open model’s weights, and the authors measured a substantial benchmark change in their test setup. It does not establish that this can be done cheaply, reliably or safely for arbitrary models. Readers should treat the figures as preliminary results until outside researchers test the method and its side effects.
Key Questions
What did GPT-6 Astra change?
According to the researchers, Astra removed a direction linked to repetitive outputs from selected weights in Qwen3.5-0.8B’s layers 15 and 19. The edit affected 131,072 weights, which the post gives as 0.016% of the model’s total.
Did the edit eliminate looping?
No. The reported share of MATH-500 answers reaching the 32,768-token limit fell from 93.0% to 33.1%, so the failure remained in a substantial portion of outputs.
What does “without training” mean here?
The authors say they did not use gradient descent, backward passes or loss-based fitting. Astra did use an iterative search with practice problems, Python code and repeated model evaluations, so the phrase does not mean the process required no testing or computational effort.
Has the result been peer-reviewed?
The source account describes the work as a blog post published November 2, 2026, and says it is not peer-reviewed. The authors also characterize the findings as very early.
Does the score show that Qwen became broadly better at math?
Not by itself. The experiment targeted a known looping problem, and the authors explicitly distinguish fixing loops from improving mathematical ability generally. Independent tests would be needed to establish how broadly the gains apply.
Source: ThorstenMeyerAI.com
